Hearing distributions • soundeR

sonify_data() plays rows in order, which works when the order means something: games in a season, sleds crossing a finish line. But a lot of data isn’t a sequence. For the ages of houses in a state, the order of the rows is arbitrary. What matters is the distribution: how many houses there are of each age.

A histogram shows a distribution by sorting values into bins and drawing a bar for each one. sonify_histogram() does the same thing for your ears. It sweeps across the bins from lowest to highest, playing one note per bin. The more rows in a bin, the higher and louder the note. Empty bins are silent.

library(soundeR)
library(dplyr)

Houses in Nebraska

ne_house_years counts the single-family houses in Nebraska by the year they were built, from parcel records. (About 13,000 parcels have no build year, so they aren’t here.)

ne_house_years
#> # A tibble: 182 × 2
#>    year_built houses
#>         <int>  <int>
#>  1       1800      4
#>  2       1820      1
#>  3       1821      1
#>  4       1840      2
#>  5       1848      1
#>  6       1849      2
#>  7       1850     10
#>  8       1851      1
#>  9       1853      1
#> 10       1854      3
#> # ℹ 172 more rows

The data is already counted, one row per year, so the counts go to weight. (With raw data, one row per house, you’d leave weight out.) Let’s start with one bin per decade:

ne_house_years |>
  sonify_histogram(year_built, weight = houses, binwidth = 10, instrument = "marimba")

The whole sweep takes 20 seconds, one note per decade. The 1800s start low and climb to a boom in the first decades of the 1900s. Then comes a sharp drop in the 1930s, the Depression, followed by a long rise after World War II that peaks in the 1970s, a dip in the 1980s, another peak in the 2000s, and a fall after the 2008 housing crash.

notes() shows each bin’s edges and count:

ne_house_years |>
  sonify_histogram(year_built, weight = houses, binwidth = 10) |>
  notes() |>
  select(bin_start, bin_end, note, value)
#> # A tibble: 21 × 4
#>    bin_start bin_end note  value
#>        <dbl>   <dbl> <chr> <dbl>
#>  1      1800    1810 C3        4
#>  2      1820    1830 C3        2
#>  3      1840    1850 C3        5
#>  4      1850    1860 C3       59
#>  5      1860    1870 C3      216
#>  6      1870    1880 C3     1066
#>  7      1880    1890 D3     3721
#>  8      1890    1900 E3     9940
#>  9      1900    1910 E4    31752
#> 10      1910    1920 A4    36364
#> # ℹ 11 more rows

Each bin includes its start but not its end, so the 1900 bin covers 1900 through 1909.

The bins change what you hear

Here’s the same data with one bin per year:

ne_house_years |>
  sonify_histogram(year_built, weight = houses, binwidth = 1, instrument = "marimba")

It sounds completely different. Through the late 1800s and early 1900s, jarring high notes jump out every five and ten years. Those are years like 1900, which has 19,205 houses, compared with 730 in 1901:

ne_house_years |>
  filter(year_built >= 1898, year_built <= 1906)
#> # A tibble: 9 × 2
#>   year_built houses
#>        <int>  <int>
#> 1       1898   1010
#> 2       1899    220
#> 3       1900  19205
#> 4       1901    730
#> 5       1902    540
#> 6       1903    816
#> 7       1904    640
#> 8       1905   4116
#> 9       1906   1608

Did Nebraska really build 26 times as many houses in 1900 as in 1901? Before 1950, 42% of the houses in this data have a year ending in 0, where you’d expect about 10%. From 1980 on, the round-number spikes mostly disappear. A pattern that regular says something about how the years got into the records, not just about when houses went up. Why it happened isn’t documented yet. It’s the kind of question to take to the people who keep the data before you report a number like “19,205 houses built in 1900.”

Neither version is wrong. Decade bins show the overall shape and hide the rounding. Year bins show the rounding and bury the shape. Picking a bin width is a choice, in sound just as in a chart, so try more than one.

With pan = TRUE, the sound also travels from your left speaker to your right as the sweep moves across the years, so you can hear where you are along the axis. It works best with headphones:

ne_house_years |>
  sonify_histogram(year_built, weight = houses, binwidth = 10, pan = TRUE, instrument = "marimba")

A few big bins can drown out the rest

When a handful of bins are far bigger than the others, most notes crowd together at the bottom of the range. log = TRUE bases pitch on the logarithm of the counts, which spreads out the small ones:

ne_house_years |>
  sonify_histogram(year_built, weight = houses, binwidth = 1, log = TRUE, instrument = "marimba")

Now you can hear the handful of houses from the early 1800s rising toward the 1900s instead of sitting on one low note. The trade-off is that differences between big bins sound smaller than they are.

A smooth curve

A histogram’s bins are all-or-nothing: a house built in 1909 counts toward one bar and a house built in 1910 toward the next. A density curve, like ggplot2::geom_density(), smooths the distribution instead. sonify_density() plays that curve as one continuous tone that glides up where there are more houses and down where there are fewer:

ne_house_years |>
  sonify_density(year_built, weight = houses, adjust = 4)

adjust sets how smooth the curve is, the same as in geom_density(). With adjust = 4 you hear three rises: to a peak around 1916, a dip in the late 1930s, a long climb to about 1971, a sag around 1992, and a last peak around 2001. With the default of adjust = 1, the curve is detailed enough to wobble at 1901, 1910 and 1920, the round-number years again:

ne_house_years |>
  sonify_density(year_built, weight = houses)

The glide uses soundeR’s built-in tones. To hear the curve on a real instrument, add glide = FALSE, which plays it as a quick run of notes:

ne_house_years |>
  sonify_density(year_built, weight = houses, adjust = 4, glide = FALSE, instrument = "marimba")

Watch the edges of your data

The last two years in this data, 2025 and 2026, were only partly recorded when it was compiled: 408 houses and 1. Played as they are, they sound like the housing market fell off a cliff. Leaving out the incomplete years is more honest:

ne_house_years |>
  filter(year_built < 2025) |>
  sonify_histogram(year_built, weight = houses, binwidth = 5, instrument = "marimba")

Raw data works too

Without weight, every row counts once. Here’s how the Husker basketball margins were spread across the 2025-26 season, in five-point bins:

husker_games |>
  sonify_histogram(point_margin, binwidth = 5, instrument = "xylophone", length = 10)

The sweep starts with the seven losses. They’re spread thin across the left side, with a silent gap where no game landed, so their notes are low: remember, pitch here is how many games fell in a bin, not how big the margin was. The notes climb for the most common results, wins by 0 to 20 points, then scatter through the blowouts, including five wins by 30 or more.

Seeing it

sonify_video() draws a sonified histogram as a histogram, with bars that fill in as the sweep passes them. A density curve fills in the same way.

ne_house_years |>
  sonify_histogram(year_built, weight = houses, binwidth = 1, instrument = "marimba") |>
  sonify_video(
    "house-years.mp4",
    title = "Houses built in Nebraska, by year",
    subtitle = "One bar per year. Listen for the round numbers.",
    x_label = "Year built",
    y_label = "Houses"
  )

Things to try

Try binwidth = 5. Do the round-number spikes go away, or just move?
Which do you trust more to describe housing in Nebraska, the year version or the decade version? What would you tell a reader about each?
Leave out the years before 1950 with filter(). How does the rest of the distribution sound without the round-number years?
Find a column in your own data where row order doesn’t matter, and listen to its distribution.